Simplifying Distributed File Systems Using a Shared Logical Disk

Abstract:

We present a new approach to building non-centralized (``serverless'')
distributed file systems, using the shared logical disk, which
provides the abstraction of a fault-tolerant, entry-consistent, shared
array of persistent data blocks. By separating the storage management
level from higher-level file system functions like metadata
management, we significantly reduce the overall complexity of the file
system's implementation.
We describe the implementation of a file system and an underlying
shared logical disk, designed to run on workstation clusters and on
the SHRIMP multicomputer. The file system level is very similar to a
conventional uniprocessor file system, relying on the underlying
shared logical disk to handle all communication, caching, and
coherence operations. The shared logical disk provides some simple
hooks to allow the file system to express its fault-tolerance
requirements. We believe the resulting design is much simpler than
other non-centralized distributed file systems.